File size: 7,831 Bytes
2d05890 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 | #!/usr/bin/env python3
"""Combine llama-server log, RSS samples, and bench results into one telemetry JSON.
Output schema (mac_bench/<chip>-<ram>gb-<size>/telemetry.json):
{
"hardware": {"chip": "M2-Max", "ram_gb": 96, "fanless": false},
"model": {"size": "2b", "repo": "continker/Qwen3.5-2B-metro-v23", "gguf_gb": 1.27},
"eval": {"tier1_composite": 84.0, "metrollm_composite": 81.5, ...},
"perf": {
"decode_tok_s_median": 41.2, "decode_tok_s_p10": 38.0, "decode_tok_s_p90": 44.5,
"decode_tok_s_n": 421,
"ttft_ms_median": 287, "ttft_ms_p90": 540,
"peak_rss_gb": 1.6,
"runner_wallclock_s": 4520
}
}
Stdin/stdout: pure JSON dump on success. Errors go to stderr; exit code is 0
unless required inputs missing.
"""
from __future__ import annotations
import argparse
import json
import re
import statistics
from pathlib import Path
# llama.cpp 'eval time' line shapes vary across versions. Cover the ones we'll see.
# Examples:
# eval time = 234.56 ms / 50 tokens ( 4.69 ms per token, 213.42 tokens per second)
# eval time = 234.56 ms / 50 runs ( 4.69 ms per token, 213.42 tokens per second)
EVAL_RE = re.compile(
r"eval time\s*=\s*([\d.]+)\s*ms\s*/\s*(\d+)\s*(?:tokens|runs)\s*"
r"\(\s*[\d.]+\s*ms per token,\s*([\d.]+)\s*tokens per second\)",
re.IGNORECASE,
)
# Some builds use 'predicted' instead of 'eval':
PRED_RE = re.compile(
r"predicted\s*=\s*([\d.]+)\s*ms\s*/\s*(\d+)\s*(?:tokens|runs)\s*"
r"\(\s*[\d.]+\s*ms per token,\s*([\d.]+)\s*tokens per second\)",
re.IGNORECASE,
)
def parse_decode_tok_s(log_path: Path) -> list[float]:
"""Parse only DECODE eval lines (skip 'prompt eval' which is ~10x faster
and would skew the median upward). The decode line is `eval time = ...`
without the 'prompt' prefix. We require at least 8 tokens evaluated to
skip 1-2 token completion bursts."""
if not log_path.exists():
return []
rates: list[float] = []
with log_path.open() as f:
for line in f:
# CRITICAL: skip prompt-eval lines (regex would match them otherwise).
if "prompt eval time" in line:
continue
for rx in (EVAL_RE, PRED_RE):
m = rx.search(line)
if m:
n_tokens = int(m.group(2))
tok_s = float(m.group(3))
if n_tokens >= 8:
rates.append(tok_s)
break
return rates
def parse_peak_rss_gb(rss_log: Path) -> float:
if not rss_log.exists():
return 0.0
peak_kb = 0
with rss_log.open() as f:
for line in f:
parts = line.split()
if len(parts) >= 2 and parts[1].isdigit():
peak_kb = max(peak_kb, int(parts[1]))
return peak_kb / 1024 / 1024 # KB → GB
def percentile(values: list[float], p: float) -> float:
if not values:
return 0.0
s = sorted(values)
idx = max(0, min(len(s) - 1, int(round((p / 100.0) * (len(s) - 1)))))
return s[idx]
def parse_runner_ttft(raw_path: Path) -> list[float]:
"""Pull TTFT (ms) from runner output's per-case latency. Different runner versions
expose this differently; we tolerate missing fields."""
if not raw_path.exists():
return []
try:
data = json.loads(raw_path.read_text())
except json.JSONDecodeError:
return []
cases = data.get("cases") or data.get("results") or []
out: list[float] = []
for c in cases:
# try common field names
for key in ("ttft_ms", "first_token_ms", "first_round_latency_ms"):
v = c.get(key)
if isinstance(v, (int, float)):
out.append(float(v))
break
else:
# fallback: nested under 'latency' or 'timing'
timing = c.get("latency") or c.get("timing") or {}
v = timing.get("ttft_ms") or timing.get("first_token_ms")
if isinstance(v, (int, float)):
out.append(float(v))
return out
def load_metrics(scored_path: Path) -> dict:
"""Pull tier1, composite, and n_cases from the scored output. Field
locations differ slightly from what the runner produces — we read both
`metrics.tier1_composite` (the leaderboard number) and
`summary.cases_scored` (the n)."""
if not scored_path.exists():
return {}
try:
d = json.loads(scored_path.read_text())
except json.JSONDecodeError:
return {}
metrics = d.get("metrics", {}) or {}
summary = d.get("summary", {}) or {}
scores = d.get("scores", []) or []
n_cases = summary.get("cases_scored") or len(scores) or None
tier1_pct_values = [s.get("tier1_pct") for s in scores if isinstance(s, dict) and s.get("tier1_pct") is not None]
tier1_pct_mean = (sum(tier1_pct_values) / len(tier1_pct_values)) if tier1_pct_values else None
return {
"tier1_composite": metrics.get("tier1_composite"),
"metrollm_composite": metrics.get("metrollm_composite"),
"tier1_pct_mean": tier1_pct_mean,
"n_cases": n_cases,
}
def fanless_for_chip(chip: str) -> bool:
# Apple silicon fanless skus: MacBook Air (M1/M2/M3 base/Pro variants don't ship fanless),
# only the **base** Air chips (M1, M2, M3, M4 Air) are fanless.
# Pro/Max/Ultra are all fan-cooled. Match conservatively.
fanless_chips = {"M1", "M2", "M3", "M4"}
base = chip.replace("-", " ").strip()
return base in fanless_chips
def main():
p = argparse.ArgumentParser()
p.add_argument("--llama-log", required=True, type=Path)
p.add_argument("--rss-log", required=True, type=Path)
p.add_argument("--raw-results", required=True, type=Path)
p.add_argument("--scored-results", required=True, type=Path)
p.add_argument("--runner-wallclock", required=True, type=int)
p.add_argument("--chip", required=True)
p.add_argument("--ram-gb", required=True, type=int)
p.add_argument("--size", required=True)
p.add_argument("--ctx-size", required=True, type=int)
p.add_argument("--output", required=True, type=Path)
args = p.parse_args()
rates = parse_decode_tok_s(args.llama_log)
ttfts = parse_runner_ttft(args.raw_results)
peak_rss = parse_peak_rss_gb(args.rss_log)
metrics = load_metrics(args.scored_results)
gguf_path = Path("data/mac_models") / f"Qwen3.5-{args.size.upper()}-metro-v23-Q4_K_M.gguf"
gguf_gb = gguf_path.stat().st_size / 1e9 if gguf_path.exists() else 0.0
out = {
"hardware": {
"chip": args.chip,
"ram_gb": args.ram_gb,
"fanless": fanless_for_chip(args.chip),
},
"model": {
"size": args.size,
"repo": f"continker/Qwen3.5-{args.size.upper()}-metro-v23",
"gguf_gb": round(gguf_gb, 3),
"ctx_size": args.ctx_size,
},
"eval": {
"tier1_composite": metrics.get("tier1_composite"),
"metrollm_composite": metrics.get("metrollm_composite"),
"tier1_pct_mean": metrics.get("tier1_pct_mean"),
"n_cases": metrics.get("n_cases"),
},
"perf": {
"decode_tok_s_median": statistics.median(rates) if rates else 0.0,
"decode_tok_s_p10": percentile(rates, 10),
"decode_tok_s_p90": percentile(rates, 90),
"decode_tok_s_n": len(rates),
"ttft_ms_median": statistics.median(ttfts) if ttfts else 0.0,
"ttft_ms_p90": percentile(ttfts, 90),
"ttft_ms_n": len(ttfts),
"peak_rss_gb": round(peak_rss, 3),
"runner_wallclock_s": args.runner_wallclock,
},
}
args.output.write_text(json.dumps(out, indent=2))
print(f"Wrote {args.output}")
if __name__ == "__main__":
main()
|